CC-DeePC: adaptive conformal calibration for safe data-enabled predictive control under non-stationary dynamics
Abstract
Data-Enabled Predictive Control (DeePC) regulates unknown systems directly from measured input–output trajectories, but its behavioral predictor is only as reliable as the data window it is built from, and its constraint set offers no protection once the underlying dynamics drift away from that window. We propose CC-DeePC, a conformally-calibrated variant of DeePC that tightens the receding-horizon output constraint using an online, distribution-free quantile from Adaptive Conformal Inference (ACI) applied to the predictor’s own one-step residuals, instead of external agents’ trajectories as in prior conformal safe-planning work. We prove a finite-horizon empirical miscoverage bound on this residual sequence under arbitrary non-stationary dynamics (Proposition 1)—a bound on residual miscoverage, not directly on physical constraint violation—and give a sufficient condition (Lemma 1) and violation-rate corollary bridging it to physical constraint safety despite the soft slack feasibility requires. Compared against four baselines (no-margin, fixed worst-case, static conformal, and a fixed distributionally-robust margin), raw violation-rate and RMSE are statistically indistinguishable from the distributionally-robust baseline under drift, but CC-DeePC grounds substantially more operation in certified, slack-free safety: its clean-safety rate is 3.6–7×\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times $$\end{document} higher than DR-DeePC’s, 1.4–1.9×\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times $$\end{document} higher than static conformal’s, and up to two orders of magnitude higher than the fixed baseline’s, confirmed significant (Holm-corrected) in 23 of 24 tests spanning single-input (n=100\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n{=}100$$\end{document}) and 2-input/2-output (n=60\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$n{=}60$$\end{document}) benchmarks. A causal-isolation ablation attributes this to the ACI recursion, not recent data alone. Two checks probe generality: refitting drift parameters to a real NGSIM I-80 vehicle trajectory reproduces the same pattern, and a harder, moving-obstacle 2D navigation extension with LiDAR sensing shows a lower collision rate than three of four baselines but worse tracking error and path length, with the clean-safety advantage largely gone once sensing is imperfect. Code, data, and protocol are released for reproducibility.